Don’t surrender data control in pursuit of intelligence

With organizations reorganizing themselves to put artificial intelligence at the center, we’ve seen a shift from information being valuable to being a means of achieving intelligence, which is far more important.

But what happens to organizations’ data when they adopt AI? Do they keep it or turn it over to AI and cloud vendors? Do they even know where their data is?

As AI moves from experimental to essential, leaders should ask themselves if they’re doing enough to control their own data.

This isn’t a philosophical question. Recent announcements from large information technology providers have raised concerns. ServiceNow Inc., for example, recently launched Context Engine, a product that aggregates customer data into a unified layer and updates it in real time. But there’s a possible catch for customers who want to use their own AI agents outside the ServiceNow ecosystem: they could incur consumption charges for usage beyond their chosen level. As Constellation Research Inc. noted in an April blog post, ServiceNow customers will see their usage-based meters increase as consumption rises.

To me, the question is about control more than cost. Your data – which is your institutional knowledge and accumulated intelligence – is at risk of becoming someone else’s asset when you enter into any agreement that places it in the center of a vendor ecosystem.

When you become dependent on a vendor, you lose some control. CIO.com quoted Info-Tech Research Group Inc. Advisory Fellow Scott Bickley, noting that Context Engine creates an “implied dependence” on ServiceNow’s data and governance models, as well as its platform architecture. That’s a valid concern.

The missing ingredient

To make full use of data by turning it into actionable intelligence, you need context. And context doesn’t come for free; it has to be engineered.

Context engineering is the discipline of structuring, curating, and governing the information on which an AI operates. It’s not enough to point a model at your data and let it run. The model needs to understand which data is authoritative, which is current, which is relevant to a given decision, and critically, what it doesn’t know. Without deliberate context engineering, even a powerful LLM will confidently hallucinate, drawing on stale, incomplete or simply wrong information to produce outputs that feel credible but aren’t.

Context engineering is driven by bringing AI tools into an organization’s data boundary, not by pushing data out to someone else’s cloud. Any AI architecture needs to be able to construct the rich contextual layer that those tools need to reason accurately. This is what is meant by data sovereignty. AI works within your context, not its own.

But context engineering alone isn’t sufficient. The other half of the equation is validation and verification of AI outputs before they become part of your organizational knowledge base. This is a step many organizations are skipping entirely, and it may be the most consequential mistake of the current AI adoption wave.

When an LLM generates an answer, a summary, a recommendation, or a report, that output needs to be tested against ground truth before it gets filed, shared, acted on, or worst of all, used to train the next model in your stack (leading to the dreaded model collapse). Once a hallucinated or subtly wrong output gets embedded in your knowledge infrastructure, it propagates and becomes the context that shapes future AI outputs.

The organizations that get this right will build AI systems that grow more accurate and more trustworthy over time. Those that don’t will find themselves managing an ever-compounding contagion liability, and one they cannot easily audit because the original error is buried under layers of downstream inference.

Critical decisions

We’re relatively early on the AI adoption and implementation curve. Organizations are making critical decisions now about the AI architecture and platforms they want to adopt. While there’s a lot of money and organizational energy on the line, the choice of an AI vendor is not permanent.

What is more lasting is the approach enterprises and governments take to subjecting their data to the whims of outside entities. Will they take every step to maintain control? Or will they cede a portion of it to someone else?

We have rapidly adopted LLMs as tools that give us intelligence, a step beyond information, which is a step ahead of the raw building blocks of data. In this stair-step progression, we should ask ourselves if we’ve become so enamored with intelligence that we’ve forgotten to protect our data. The answer might prove uncomfortable.

Richard Boyd is CEO and co-founder of UltiSim Inc., a Chapel Hill, North Carolina-based enterprise AI infrastructure and digital twin company. He wrote this article for SiliconANGLE.

Image: ChatGPT/SiliconANGLE